Skip to content
Python

fix DeepSeek V4 context windows

Amar Teja Kommineni9ae0f36bf2eb
diff --git a/coworker/providers/matrix.py b/coworker/providers/matrix.py
index 6dcee64b..0c15d4a6 100644
--- a/coworker/providers/matrix.py
+++ b/coworker/providers/matrix.py
@@ -187,11 +187,13 @@ MATRIX: dict[str, ModelEntry] = {
# GLM-5.2 shipped with a 1M window (docs.z.ai, read 2026-09-08: "1M" context,
# 128K output); the 128,000 here was stale and compacted it at 102,400 (OPE-170).
"zai:glm-5.2": ModelEntry("GLM-5.2 · Z AI", _AGENTIC, 1_000_000),
+ # DeepSeek's native API exposes the full 1M context for both V4 variants
+ # (verified against the official model table 2026-08-10).
"deepseek:deepseek-v4-flash": ModelEntry(
- "DeepSeek V4 Flash · DeepSeek", _AGENTIC, 128_000
+ "DeepSeek V4 Flash · DeepSeek", _AGENTIC, 1_000_000
),
"deepseek:deepseek-v4-pro": ModelEntry(
- "DeepSeek V4 Pro · DeepSeek", _AGENTIC, 128_000
+ "DeepSeek V4 Pro · DeepSeek", _AGENTIC, 1_000_000
),
"kimi:kimi-k2.6": ModelEntry("Kimi K2.6 · Moonshot", _AGENTIC, 256_000),
"minimax:MiniMax-M2.5": ModelEntry("MiniMax M2.5 · MiniMax"),
@@ -223,6 +225,7 @@ MATRIX: dict[str, ModelEntry] = {
"Kimi K2.6 · via Together", _AGENTIC, 262_144
),
"together:deepseek-ai/DeepSeek-V4-Pro": ModelEntry(
+ # Together's serverless deployment exposes 512K, not the model-level 1M.
"DeepSeek V4 Pro · via Together", _AGENTIC, 512_000
),
"together:meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8": ModelEntry(
@@ -236,7 +239,7 @@ MATRIX: dict[str, ModelEntry] = {
"Kimi K2.6 · via Fireworks", _AGENTIC, 256_000
),
"fireworks:accounts/fireworks/models/deepseek-v4-pro": ModelEntry(
- "DeepSeek V4 Pro · via Fireworks", _AGENTIC, 128_000
+ "DeepSeek V4 Pro · via Fireworks", _AGENTIC, 1_048_576
),
"fireworks:accounts/fireworks/models/llama4-maverick-instruct-basic": ModelEntry(
"Llama 4 Maverick · via Fireworks", _AGENTIC, 1_000_000
@@ -261,7 +264,7 @@ MATRIX: dict[str, ModelEntry] = {
"Kimi K2.6 · via OpenRouter", _AGENTIC, 256_000
),
"openrouter:deepseek/deepseek-v4-pro": ModelEntry(
- "DeepSeek V4 Pro · via OpenRouter", _AGENTIC, 128_000
+ "DeepSeek V4 Pro · via OpenRouter", _AGENTIC, 1_048_576
),
"openrouter:meta-llama/llama-4-maverick": ModelEntry(
"Llama 4 Maverick · via OpenRouter", _AGENTIC, 1_000_000
diff --git a/tests/test_token_usage.py b/tests/test_token_usage.py
index 663bf7f1..9951b523 100644
--- a/tests/test_token_usage.py
+++ b/tests/test_token_usage.py
@@ -360,3 +360,16 @@ def test_model_context_windows_covers_verified_entries_only():
assert windows["anthropic:claude-fable-5"] == 1_000_000
assert "together:thinkingmachines/Inkling" not in windows # unverified stays absent
assert all(isinstance(v, int) and v > 0 for v in windows.values())
+
+
+def test_deepseek_v4_context_windows_match_provider_limits():
+ windows = model_context_windows()
+ expected = {
+ "deepseek:deepseek-v4-flash": 1_000_000,
+ "deepseek:deepseek-v4-pro": 1_000_000,
+ "together:deepseek-ai/DeepSeek-V4-Pro": 512_000,
+ "fireworks:accounts/fireworks/models/deepseek-v4-pro": 1_048_576,
+ "openrouter:deepseek/deepseek-v4-pro": 1_048_576,
+ }
+
+ assert {model: windows[model] for model in expected} == expected